Add aten::soft_margin_loss and its variants

[yucai-intel]

soft_margin_loss soft_margin_loss_backward

[xytintel]

Close this PR because CUDA is not registered separately; it is a composite op.